Unità di ricerca "Artificial Intelligence-based eHealth"

L’Intelligenza Artificiale (IA) è sempre più pervasiva in vari ambiti della nostra vita. La medicina, e più in particolare la telemedicina, non fanno eccezione: è opinione unanime, infatti, che siamo alle soglie di una rivoluzione nell’assistenza sanitaria. Le tecnologie digitali ci permettono oggi di generare una grande quantità di dati; la sfida è utilizzare tale mole di dati per generare nuova conoscenza clinicamente rilevante. L’IA sviluppa algoritmi di varia natura al fine di imparare a comprendere i dati a disposizione.

In particolare, tecniche proprie dell’IA – quali l’apprendimento automatico, la visione artificiale, la programmazione basata su regole, il pattern recognition e il deep learning, ecc. –, sono oggi in grado di identificare pattern significativi nei dati acquisiti e possono essere pertanto applicate efficacemente come ausilio a vari compiti.

Per esempio, possono migliorare la precisione di diagnosi mediche, facilitare lo sviluppo di farmaci, aiutare a prendere decisioni sui trattamenti da somministrare ai pazienti, simulare possibili interventi terapeutici e anticiparne l’esito, come pure ottimizzare processi, decisioni operative e utilizzo delle risorse finanziarie proprie dell’ecosistema sanitario.

Gli operatori sanitari potranno trarre beneficio dall’integrazione dell’IA nella loro pratica clinica tradizionale al fine di affrontare problemi che, per loro complessità, si rivelano onerosi in termini di tempo e, spesso, proni a inefficienza. In questo contesto, medici e operatori sanitari, con la loro capacità di valutare criticamente e utilizzare a proprio vantaggio le predizioni prodotte dagli algoritmi di IA, assumono un ruolo centrale nei processi decisionali diagnostici e terapeutici.

Consigliere Citel

semeraro__834x834.jpg

Il team

castellano.jpg
damato.jpg
degemmis.jpg
lops.jpg
basile.jpg
vessio.jpg
decarolispepper.jpg
stefano-ferilli.png
polignano.png
musto.jpg

Pubblicazioni

Image
A deep learning model for the analysis of medical reports in ICD-10 clinical coding task. To appear
Polignano, M., Basile, P., De Gemmis, M., Lops, P. & Semeraro, G. (2020)
The practice of assigning a uniquely identifiable and easily traceable code to pathology from medical diagnoses is an added value to the current modality of archiving health data collected to build the clinical history of each of us. Unfortunately, the enormous amount of possible pathologies and medical conditions has led to the realization of extremely wide international codifications that are difficult to consult even for a human being. This difficulty makes the practice of annotation of diagnoses with ICD-10 codes very rare. In order to support this operation, a classification model was proposed, able to analyze medical diagnoses written in natural language and automatically assign one or more international reference codes. The model has been evaluated on a dataset released in the Spanish language for the eHealth challenge (CodiEsp) of the international conference CLEF 2020, but it could be extended to any language with Latin characters. Although still far from an accuracy sufficient to do without a medical opinion, the results obtained show the feasibility of the task and are a starting point for future studies in this direction.
Image
HealthAssistantBot: A Personal Health Assistant for the Italian Language. IEEE Access, 8, 107479-107497
Polignano, M., Narducci, F., Iovine, A., Musto, C., De Gemmis, M., & Semeraro, G. (2020)
In this article, we present HealthAssistantBot, an intelligent virtual assistant able to talk with patients in order to understand their symptomatology, suggest doctors, and monitor treatments and health parameters. In a simple way, by exploiting a natural language-based interaction, the system allows the user to create her health profile, to describe her symptoms, to search for doctors or to simply remember a treatment to follow. Specifically, our methodology exploits machine learning techniques to process users symptoms and to automatically infer her diseases. Next, the information obtained is used by our recommendation algorithm to identify the nearest doctor who can best treat the user’s condition, considering the community data. In the experimental session we evaluated our HealthAssistantBot with both an offline and online evaluation. In the first case, we assessed the performance of our internal components, while in the second one we carried out a study involving 102 subjects who interacted with the conversational agent in a daily use scenario. Results are encouraging and showed the effectiveness of the strategy in supporting the patients in taking care of their health.
Image
Power to the patients: The HealthNetsocial network. Information Systems, 71, 111-122
Narducci, F., Lops, P., & Semeraro, G. (2017)
HealthNet (HN) is a social network that brings together patients with similar health conditions. HN helps users in finding a solution to their health problems by suggesting doctors and health facilities that best fit the patient profile. Indeed, the core component of HN is a recommender system that suggests patients similar to the target user and supports the choice of the doctor and the hospital for a specific condition. The recommendation algorithm first computes similarities among patients, and then generates a ranked list of doctors and hospitals for a given patient profile by exploiting health data shared by the community. The HN typical user can find the most similar patients, can look how they treated their diseases, and can receive suggestions for solving her condition. In order to facilitate the interaction with the system and improve the recommendation step, the patient can express her health status by a natural-language sentence. The system analyzes the sentence and identifies the most relevant medical area (e.g., orthopedics, neurology, allergology, etc.) for that specific case, and uses this information for the recommendation task. Currently HN is in alpha version and only for Italian users, but in the future we want to extend the platform to other languages. We carried out both an in-vitro experimental evaluation to assess the effectiveness of the module for analyzing natural language descriptions provided by users as well as the recommender system to suggest the right doctors for a specific health problem, and an in-vivo evaluation performed by real doctors. Results are really encouraging.
Image
Recommending doctors and health facilities in the HealthNet Social Network. In Proceedings of the Workshop on Social Media for Personalization and Search (SoMePeAS 2017), co-located with 39th European Conference on Information Retrieval (ECIR 2017), CEUR-WS vol. 1840 (pp.1-7).
Narducci, F., Musto, C., Polignano, M., De Gemmis, M., Lops, P. & Semeraro, G. (2017)
In this paper we present HealthNet (HN), a social network that helps patients to meet the best doctor for her health condition. The core component of HN is a recommender system that suggests to the user patients similar to her, and generates suggestions about doctors and hospitals that best match her patient profile. Currently an alpha version of HN is available only for Italian users, but in the next future we want to extend the platform to other languages. We organized three focus groups with patients, practitioners, and health organizations in order to obtain comments and suggestions. All users were very enthusiastic by using the prototype version of HN1.
Image
A recommender system for connecting patients to the right doctors in the healthnet social network. In Proceedings of the 24th international conference on World Wide Web (pp. 81-82)
Narducci, F., Musto, C., Polignano, M., de Gemmis, M., Lops, P., & Semeraro, G. (2015, May)
In this work we present a semantic recommender system able to suggest doctors and hospitals that best fit a specific patient profile. The recommender system is the core component of the social network named HealthNet (HN). The recommendation algorithm first computes similarities among patients, and then generates a ranked list of doctors and hospitals suitable for a given patient profile, by exploiting health data shared by the community. Accordingly, the HN user can find her most similar patients, look how they cured their diseases, and receive suggestions for solving her problem. Currently, the alpha version of HN is available only for Italian users, but in the next future we want to extend the platform to other languages. We organized three focus groups with patients, practitioners, and health organizations in order to obtain comments and suggestions. All of them proved to be very enthusiastic by using the HN platform.
Image
A recommender system for connecting patients to the right doctors in the healthnet social Contact-less real-time monitoring of cardiovascular risk using video imaging and fuzzy inference rules. Information, 10(1), 9
Casalino, G., Castellano, G., Pasquadibisceglie, V., & Zaza, G. (2019) DOI: 10.3390/info10010009
In this work, we propose an innovative solution for monitoring cardiovascular parameters that is low cost and can be easily integrated within any common home environment. The proposed system is a contact-less device composed of a see-through mirror equipped with a camera that detects the person’s face and processes video frames using photoplethysmography in order to estimate the heart rate, the breath rate and the blood oxygen saturation. In addition, the color of lips is automatically detected via clustering-based color quantization. The estimated parameters are used to predict a risk of cardiovascular disease by means of fuzzy inference rules integrated in the mirror-based monitoring system. Comparing our system to a contact device in measuring vital parameters on still or slightly moving subjects, we achieve measurement errors that are within acceptable margins according to the literature. Moreover, in most cases, the response of the fuzzy rule-based system is comparable with that of the clinician in assessing a risk level of cardiovascular disease.
Image
A Predictive Model for MicroRNA Expressions in Pediatric Multiple Sclerosis Detection. Modeling Decisions for Artificial Intelligence. Lecture Notes in Computer Science, LNAI 11676, 177–188.
Casalino, G.,Castellano, G., Consiglio, A., Liguori, M., Nuzziello, N., & Primiceri, D. (2019) DOI: 10.1007/978-3-030-26773-5_16
The study of miRNA data can be of valuable support for the early diagnosis of multifactorial diseases such as pediatric Multiple Sclerosis. However the analysis of miRNA expressions poses several challenges due to high dimensionality and imbalance of data. In this paper we present a data science workflow to develop a predictive model that is intended to support the clinicians in the diagnosis of Multiple Sclerosis starting from miRNA data produced by Next-Generation Sequencing. The goal is to create an effective model able to predict the pathological condition of a patient starting from his miRNA expression profile. Based on the proposed workflow, the miRNA dataset is firstly preprocessed in order to reduce its high dimensionality (from 1287 features to 40 features) and to mitigate class imbalance. Then a classification model is learnt from data via neural network training. Results show that the model defined by using the 40 data-driven selected features achieves an overall classification accuracy of 94% on test data and overcomes the model based on 42 features selected by the experts that achieves only 83% of overall accuracy.
Image
Dynamic incremental semi-supervised fuzzy c-means for bipolar disorder episode prediction. In Proc. of 23rd International Conference on Discovery Science (DS 2020). To appear
G. Casalino, G., Castellano, G., F. Galetta, F., & K. Kaczmarek-Majer, K. (2020)
Bipolar Disorder (BD) is a chronic mental illness character-ized by changing episodes from euthymia (healthy state) through depression and mania to the mixed states. In this context, data collected through the interaction of patients with smartphones enable the creation of predictive models to support the early prediction of a starting episode. Previous research on predicting a new BD episode use mostly supervised learning methods that require labeled data and hence force a filtering of the available data to retain only those data that have valid labels (from the psychiatric assessment). To avoid limitations of supervised learning, in this paper we investigate the use of a semi-supervised learning approach that combines both labeled and unlabeled data to derive a model for BD episode prediction. Specifically we apply the DISSFCM (Dynamic Incremental Semi-Supervised Fuzzy C-Means) algorithm which offers the possibility to process in an incremental fashion the data stream of the voice signal captured by the smartphone, thus exploiting the evolving time structure of data which is ignored by static learning methods. Preliminary results on real-world data collected at the Department of Affective Disorders, Institute of Psychiatry and Neurology in Warsaw (Poland) show that DISSFCM is able to predict some of healthy episodes (euthymia) and disease episodes even when only 25% of labeled data are available.
Image
A study of Machine Learning models for Clinical Coding of Medical Reports at CodiEsp 2020. Working Notes of CLEF 2020 - Conference and Labs of the Evaluation Forum.
Polignano, M., Suriano, V., Lops, P., De Gemmis, M., & Semeraro, G. (2020)
Abstract. The task of identifying one or more diseases associated with a patient’s clinical condition is often very complex, even for doctors and specialists. This process is usually time-consuming and has to take into account different aspects of what has occurred, including symptoms elicited and previous healthcare situations. The medical diagnosis is often provided to patients in the form of written paper without any correlation with a national or international standard. Even if the WHO (World Health Organization) released the ICD10 international glossary of diseases, almost no doctor has enough time to manually associate the patient’s clinical history with international codes. The CodiEsp task at CLEF 2020 addressed this issue by proposing the development of an automatic system to deal with this task. Our solution investigated different machine learning strategies in order to identify an approach to face that challenge. The main outcomes of the experiments showed that a strategy based on BERT for pre-filtering and one based on BiLSTM-CNN-SelfAttention for classification provide valuable results. We carried out several experiments on a subset of the training set for tuning the final model submitted to the challenge. In particular, we analyzed the impact of the algorithm, the input encoding strategy, and the thresholds for multi-label classification. A set of experiments has been carried out also during a post hoc analysis. The experiments confirmed that the strategy submitted to the CodiEsp task is the best performing one among those evaluated, and it allowed us to obtain a final mean average error value on the test set equal to 0.202. To support future developments of the proposed approach and the replicability of the experiments we decided to make the source code publicly accessible.